v0.3.0

0.3.0 (2026-09-13)

Breaking

• one rename tool. metadata_rename field=tags|genres|narrators|authors|languages|publishers
  replaces server_tag_rename, narrator_edit and audit_terminology_rename, and with
  remove drops a tag, genre, narrator, language or publisher everywhere. audit_terminology
  is now audit_spelling, and every group it reports is fixed by the one tool. 88 tools -> 86
• audit_no_episodes and audit_stale_feed are audit_podcast_no_episodes and
  audit_podcast_stale_feed, so the name says they only look at podcasts
• item_cover_edit no longer removes the cover when called with neither url nor file:
  pass remove=true. A call that forgot its url used to delete the cover
• the README token table is now what the model sees. Most clients send only the description
  and input schema to the model, not the output schema, which was 60% of the earlier figure

Added

• audit_unembedded: books whose audio files carry no tags, or tags that disagree with the
  current title, author, narrator, series, genres, year or publisher - what
  item_embed_metadata is due for after a curation pass. It reads the expanded items in
  batches of 50, so it runs apart from audit_all

Fixed

• item_edit clear for narrators, series, genres and tags was a silent no-op: the
  empty list was dropped from the request (omitempty), so the server saw nothing to change.
  lib/abs list fields are now omitzero: nil leaves a field alone, an empty slice clears it
• audit_cover_ratio measured the server's 400-pixel-wide cache copy rather than the cover
  file, so every cover was "400 wide" and the too-small check could never fire. It now asks for
  the raw file, reads only the image header instead of buffering the whole image, and skips
  the request for items the listing already says have no cover
• audit_spelling (then audit_terminology) never found narrator spellings: the sweep sees the minified item shape,
  which carries narrators only as one joined narratorName, and only the expanded
  narrators list was read
• an audit with a server-side filter (audit_missing, audit_issues, audit_no_audio)
  overran limit once an earlier library had filled it: the request for the next library was
  sent with no limit, which the server reads as everything
• lib/abs DeleteAuthorImage returned an empty record; the server answers with the author
  under an author key, like the image and match routes

Changed

• docker-free tests for the tools: an in-memory MCP session against a canned Audiobookshelf
  (tools/handlers_test.go), covering what the live fixtures cannot reach - a library with
  covers, inconsistent narrator spellings, more findings than the limit
• lib/abs request and response-shape tests without a server: paging, streaming, multipart
  upload, the vocabulary path encoding, the chapter and feed-episode unwrapping
• intParam/intQuery and narratorID/vocabularyID were the same function twice